On pavement crack detection using DAUNet neural network

This work is devoted to road pavement cracks segmentation problem. Proposed a Unet-like network with a pre-trained ResNet18 as an encoder (without two last classification layers) and a standard decoder is proposed. It is shown that this architecture, together with a strong augmentation system and a well-chosen loss function, is capable of surpassing state-of-the-art solutions in terms of various quality metrics.